Content-Based Visual Information Retrieval for Real-World Media Archives

Nowadays media archives, such as housed by the National Institute for Sound and Video (NISV), the British Broadcasting Corporation (BBC) and France Télévision, are huge and growing steadily. Indexing and delivering the information gets increasingly difficult. Current manual techniques of metadata annotation that are often still in use by these institutions are not future-proof.

We therefore investigate in fast, automatic indexing, querying- and deliver strategies for massive visual information (images, video). Our system is based on an interdisciplinary project from „ICT with Industry 2013“, combining viewpoints from Computer Vision, High-Performance Computing and Information Retrieval.

The core of our approach is the use of workflow frameworks, such as VS-VLAM, for modelling the system in a flexible manner. This allows us to use different Visual Object Classification methods (Bag-of-Words, Convolutional Neural Networks) based of the chosen trade-off between speed and index/query accuracy.

Publications:

• C. Kehl and A. L. Varbanescu, „Towards Distributed, Semi-Automatic Content-Based Visual Information Retrieval (CBVIR) of Massive Media Archives,“ in International Supercomputing Conference (ISC), 2015.
  [Bibtex]

  @Conference{Kehl2015a,
  Title = {{Towards Distributed, Semi-Automatic Content-Based Visual Information Retrieval (CBVIR) of Massive Media Archives}},
  Author = {Kehl, Christian and Varbanescu, Ana Lucia},
  Booktitle = {{International Supercomputing Conference (ISC)}},
  Year = {2015},
  Note = {http://www.isc-events.com/isc15_ap/auftritt/daten/attachments/12_Poster_Kehl.pdf},
  Institution = {University of Amsterdam},
  Owner = {christian},
  Pdf = {http://christian.kehl-foto.de/ISC2015_KehlVarbanescu.pdf},
  Timestamp = {2015.08.27},
  Url = {http://christian.kehl-foto.de/ISC_2015.pdf}
  }